UnifyBench++: Personalized Multimodal Benchmark

• UnifyBench++ is a unified benchmark for personalized multimodal evaluation that integrates dense textual descriptions and richer user contexts.
• It extends the UnifyBench paradigm by combining text-only and vision-plus-text inputs with model-synthesized images to test inference and controlled generation.
• The benchmark supports multiple subtasks—including recognition, reasoning, and dense generation—evaluated using both lexical metrics and semantic scoring methods.

UnifyBench++ is a unified benchmark for personalized multimodal evaluation introduced in the context of personalized reasoning for Unified Multimodal Models. It extends UnifyBench by introducing denser textual descriptions and richer user contexts, with the stated goal of bridging the gap between simple personalization benchmarks and real-world complexity. In the formulation described for Uni-Synergy, UnifyBench++ is built on top of UnifyBench’s concept-token setting, combines text-only and vision-plus-text inputs with model-synthesized images at $512\times512$, and emphasizes tightly coupled personalized understanding and personalized generation under held-out contextual information (Shen et al., 11 May 2026).

1. Design rationale and benchmark scope

The benchmark is motivated by a specific deficiency in earlier personalization evaluations. Prior work is described as having focused on “basic recognition, VQA, QA and Pure Generation,” while lacking “high-density language contexts and reasoning-intensive generation tests.” UnifyBench++ is therefore framed as a benchmark in which models must perform inferential logic before generation, rather than merely react to sparse prompts or memorize narrow concept associations (Shen et al., 11 May 2026).

A central design principle is the introduction of “denser textual descriptions” and “richer user contexts.” In the benchmark description, these contexts are expressed through heterogeneous attribute statements and reasoning prompts associated with each concept entry. This shifts evaluation away from surface-level subject-driven synthesis toward settings where the model must recover held-out attributes and then use them correctly in downstream generation. The benchmark is accordingly described as stressing both personalized understanding and personalized generation in tightly coupled tasks (Shen et al., 11 May 2026).

The benchmark’s personalization substrate is inherited from UnifyBench. It is built on top of concept tokens and approximately $1\,\text{k}$ images per concept. This suggests that UnifyBench++ preserves the token-based personalization setting while increasing contextual density and reasoning load rather than replacing the underlying subject-token paradigm.

2. Data model, modalities, and held-out context

UnifyBench++ is described as multimodal at the level of both inputs and evaluation targets. The listed modalities are text only, realized as “extra_info blocks + reasoning prompts”; vision plus text, realized as “reference images + image prompts”; and generation outputs in the form of model-synthesized images at $512\times512$ (Shen et al., 11 May 2026).

Each concept entry is associated with four components: a set of reference images, a canonical text prompt denoted as IP or TP, multiple pieces of supplementary context called “extra_info,” and paired reasoning prompts. The supplementary context is heterogeneous and may take the form of attribute sentences or short paragraphs. The reasoning prompts are natural-language questions or incomplete captions that require the extraction of the correct attribute from the available context (Shen et al., 11 May 2026).

The benchmark description also specifies construction constraints. Each piece of extra_info is paired with exactly one reasoning prompt. During training, the model sees only a subset of extra_info and prompts, described as italicized in the dataset snapshot. During evaluation, it must infer from held-out extra_info and produce correct textual reasoning. The reported split logic therefore keeps the concept set fixed while separating extra_info and prompt subsets across train, validation, and test (Shen et al., 11 May 2026).

The paper provides a concrete example centered on the concept token $\langle \text{sks} \rangle$. A training extra_info sentence states, “He usually wears a red baseball hat.” The evaluation reasoning prompt asks, “What color baseball hat does $\langle \text{sks} \rangle$ wear?” A dense generation prompt then composes multiple attributes: “A photograph of $\langle \text{sks} \rangle$ standing next to a lamppost, wearing his favorite red baseball hat, casual jeans, and a green scarf.” This example illustrates the benchmark’s intended progression from stored context, to inference, to semantically dense generation (Shen et al., 11 May 2026).

The specialized subsets are all derived from held-out splits of the same concept entries. They are named Dense Reasoning, Reasoning, Dense Generation, Reasoning Generation, and Dense Reasoning Generation. The benchmark summary additionally states that extra_info blocks per concept are typically 3–5, with 10–50 tokens per extra_info depending on density, simple IP/TP prompts of roughly 8–12 tokens, dense prompts of roughly 25–40 tokens, and reasoning chains of roughly 15–30 tokens, while also noting that exact counts per split and the total number of concepts are not specified (Shen et al., 11 May 2026).

3. Task organization and benchmark semantics

The benchmark description states that UnifyBench++ supports “seven core tasks,” indexed by whether they test understanding or generation and by density or reasoning requirement. However, the accompanying enumeration contains nine task labels rather than seven: Recognition, Visual Question Answering, Question Answering, Reasoning, Dense Reasoning, Pure Generation, Dense Generation, Reasoning Generation, and Dense Reasoning Generation. This discrepancy is part of the source description and is relevant to interpreting the benchmark taxonomy (Shen et al., 11 May 2026).

The task families are organized as follows:

Task label	Task family	Metric or formal note
Recognition (Rec.)	Personalized Understanding	“Weight”
Visual Question Answering (VQA)	Personalized Understanding	not separately formalized
Question Answering (QA)	Personalized Understanding	not separately formalized
Reasoning (Rea.)	Personalized Understanding	maximize $P(y \mid q)$
Dense Reasoning (Dense Rea.)	Personalized Understanding	like Rea. with multi-sentence extra_info
Pure Generation (Pure Gen.)	Context-Guided Generation	CLIP-T / CLIP-I
Dense Generation (Dense Gen.)	Context-Guided Generation	high-density prompts
Reasoning Generation (Rea Gen.)	Context-Guided Generation	$CP_{\text{txt}} = IP \oplus IR_{\text{txt}}$, $CP_{\text{vis}} = BP \oplus IR_{\text{vis}}$
Dense Reasoning Generation (Dense Rea Gen.)	Context-Guided Generation	infer $IR$, then generate under dense $1\,\text{k}$0

Within personalized understanding, Recognition is defined as identity classification given reference image(s) and a user token. VQA and QA evaluate attribute answering from image-plus-question or text-plus-token inputs. Reasoning requires the model to generate a free-form natural-language rationale, described as a chain-of-thought, that justifies an attribute; its formal statement is that the model outputs a sequence $1\,\text{k}$1 given extra_info and prompt $1\,\text{k}$2, and maximizes $1\,\text{k}$3. Dense Reasoning retains the same structure while increasing the informational density of extra_info blocks (Shen et al., 11 May 2026).

Within context-guided generation, Pure Generation is the standard subject-driven synthesis setting, exemplified by prompts such as “a photo of $1\,\text{k}$4.” Dense Generation increases prompt density by combining many attributes in a single specification. Reasoning Generation is explicitly compositional: it uses a compound prompt formed as $1\,\text{k}$5 or $1\,\text{k}$6, and the description states that the model should maximize $1\,\text{k}$7. Dense Reasoning Generation is the full two-phase loop in which the model first infers the relevant attribute representation $1\,\text{k}$8 from held-out extra_info and then generates under a dense compound prompt $1\,\text{k}$9 (Shen et al., 11 May 2026).

This task design makes the benchmark more than a collection of isolated subtasks. The intended evaluation object is the transition from latent concept knowledge to explicit reasoning and then to semantically constrained synthesis.

4. Evaluation protocol, metrics, and reward structure

The evaluation protocol is split between understanding metrics, generation metrics, and a reward ensemble used during reinforcement learning rather than final evaluation. For understanding, the listed metrics are BLEU between generated rationale and ground-truth extra_info, and GPT-4o-based semantic scoring in which GPT-4o judges logical accuracy. The BLEU expression is given as

$512\times512$0

For generation, the listed metrics are CLIP-T, CLIP-I, and GPT-4o dense alignment. CLIP-T is defined as

$512\times512$1

and CLIP-I as

$512\times512$2

GPT-4o dense alignment segments the prompt by punctuation, verifies each segment’s presence, and averages validity over $512\times512$3 (Shen et al., 11 May 2026).

The benchmark description also includes a reward ensemble for RL training:

$512\times512$4

Its components are described as follows. TIER is the Euclidean distance between a cosine-similarity vector and one-hot ground truth, normalized to $512\times512$5. BER is BLIP-2 cross-modal cosine similarity. DER is DINOv2 identity-preservation cosine similarity. FER is FaceNet facial-feature cosine similarity. The paper explicitly distinguishes this ensemble from final benchmark evaluation (Shen et al., 11 May 2026).

Two further evaluation properties are structurally important. First, held-out extra_info means that success on reasoning or reasoning-conditioned generation cannot be reduced to simple memorization of training-side attribute text. Second, the use of both conventional similarity metrics and GPT-4o-based judges indicates a mixed evaluation philosophy: lexical or embedding-based measures are retained, but semantic alignment under dense prompts is treated as an independent target.

5. Relation to UnifyBench, OmniPBench, and video-centric extensions

UnifyBench++ is most directly positioned as an extension of UnifyBench. In the Uni-Synergy description, UnifyBench is characterized as covering basic recognition, VQA, QA, and Pure Generation, with comparatively minimal attribute density and no explicit reasoning-generation loop. UnifyBench++ adds denser language, richer user contexts, held-out extra_info, and explicit coupling between attribute inference and downstream generation (Shen et al., 11 May 2026).

A closely related benchmark in the same research line is OmniPBench, introduced as an extension of the public UnifyBench concept set with personalized editing tasks and cross-task evaluation protocols integrating understanding, generation, and editing (Zhong et al., 11 Jan 2026). OmniPBench uses 20 concepts, 10 reference images per concept, prompt sets of approximately 50 understanding and 50 generation templates per concept, and a personalized editing suite with five editing instructions per concept. It also defines a composite benchmark score over understanding, generation, Personalized Attribute-Reasoning Generation, and editing. This neighboring benchmark is not identical to UnifyBench++, but it shows how the UnifyBench lineage has been extended toward chained, cross-task personalization protocols (Zhong et al., 11 Jan 2026).

The name “UnifyBench++” is also used in a prospective, video-centric sense in work on UniVBench. There, a “next-generation ‘UnifyBench++’” is described as a benchmark that would build on four design goals: unified coverage of perception and generation, rich multi-shot cinematic complexity, standardized agentic evaluation, and fine-grained, multi-dimensional metrics (Wei et al., 25 Feb 2026). The same source proposes scaling to thousands of videos, adding tasks such as video-question answering and cross-video retrieval, incorporating multilingual scripts and audio evaluation, and preserving an agentic, shot-level decomposition and multi-dimensional scoring architecture. In that usage, “UnifyBench++” functions as a forward-looking design template rather than the personalized concept benchmark introduced in Uni-Synergy (Wei et al., 25 Feb 2026).

A common source of confusion is therefore terminological. In the personalized reasoning literature, UnifyBench++ denotes a benchmark with held-out textual context and reasoning-conditioned image generation. In the video foundation model literature, the same name is used aspirationally to describe an expanded unified video benchmark. The overlap is conceptual rather than identical.

6. Significance, interpretive cautions, and documentation boundaries

Within its stated scope, UnifyBench++ is designed to evaluate whether a model can move from user-specific context to explicit inference and then to controlled generation. The benchmark summary characterizes this as demanding “explicit attribute reasoning from held-out context,” testing generation under both simple and dense attribute conditions, and measuring performance with both traditional metrics and advanced semantic judges. On that basis, the source concludes that it is suited to research on tightly integrated personalized understanding and generation (Shen et al., 11 May 2026).

Several interpretive cautions follow directly from the available description. Exact benchmark cardinalities are not fully specified: the number of concepts is not given, exact counts per split are not given, and the statistical summary marks several quantities as typical rather than fixed. The task inventory itself contains an internal inconsistency, since the text claims seven core tasks while enumerating nine labels. These are not minor editorial details; they delimit what can be asserted precisely about the benchmark as documented (Shen et al., 11 May 2026).

A further implication is methodological rather than numerical. Because the benchmark keeps the concept set fixed while withholding portions of extra_info and prompt subsets, success appears to depend on whether a model can internalize or retrieve concept knowledge in a way that supports inference under distributionally shifted context. This suggests that UnifyBench++ is less a pure subject-driven generation benchmark than a benchmark for reasoning-conditioned personalization.

In the broader benchmark landscape, UnifyBench++ occupies the space between classic personalization suites and unified multimodal evaluation frameworks. It preserves the concept-token paradigm of UnifyBench, anticipates cross-task chaining later made explicit in OmniPBench, and resonates with the broader push toward unified evaluation seen in video-centric work such as UniVBench. Its distinctive feature, however, is the use of held-out contextual knowledge to test whether personalized understanding and personalized generation are actually coupled rather than merely co-located within a single model (Shen et al., 11 May 2026).